OpenAI Uses 1979 MIT Experiment to Introduce Astra
OpenAI has released a teaser for its yet-to-be-launched flagship model Astra, featuring footage from the 1979 MIT experiment "Put That There," which was presented at SIGGRAPH in 1980. The footage is credited to the Media Lab, Chris Schmandt, and Eric Hulteen. The teaser does not provide product specifications or a launch date.
The experiment was completed by the Architectural Machine Group, the predecessor of the Media Lab, with related spatial data work funded by the Defense Advanced Research Projects Agency. In the video, the operator sits in front of a large screen wearing a position sensor and says "put that there," while the system combines the spoken pronouns with the finger's location to move graphics; subsequent demonstrations can create, copy, and schedule ships on a Caribbean map. The speech recognition vocabulary is around a hundred words, with grammar tailored to the scene and requiring pre-recorded speaker voices. The first author of the paper is Richard A. Bolt.
The research's starting point is not achieving 100% accuracy in speech, but recognizing that speech will never be perfect, thus requiring additional channels, syntactic semantics, context, and immediate feedback to improve "effective accuracy." When instructions are ambiguous, the machine will ask "which one," "where," or "relative to what." Gestures make pronouns usable, while speech provides precision. This is a frequently cited prototype of multimodal, referential dialogue in the history of human-computer interaction.
Astra is currently in a scaling and compensation period: paid users who cannot use the new model on the same day will receive a storage reset, as the company has slowed down the release due to network capacity assessments. Placing the 1979 laboratory footage at the beginning frames the next-generation model as one that "understands incomplete sentences," rather than another keyboard dialogue box. The experimental hardware consisted of dedicated sensors and a fixed graphical world; today, achieving the same task in a cluttered workspace with long tasks requires generality, not just slogans.
In market mechanisms, what is bought is the imagined premium of a multimodal agent, while what is sold is the yet-to-be-seen Astra interaction manual. Attention shifts from failures and limits to a segment of public domain laboratory footage. The beneficiaries are model companies that need to redefine product categories; the pressured are subscription users still measured by chat boxes. Tribute does not replace specifications; the opening cannot serve as a release note.
Supplement: The paper clearly states that combining speech with simultaneous pointing makes pronouns natural; pointing with speech makes reference precise. Without either, the phrase "put that there" does not hold.
Source: Public Information
ABAB AI Insight
Half a century ago, the laboratory wrote the failure of speech recognition as a design premise; today, model companies write it as a release aesthetic. Bolt assumed recognition would never be accurate, so machines should ask follow-up questions; OpenAI assumes models are accurate enough, so people should be encouraged to speak with their fingers again. The same English phrase, in 1979 relied on magnetic sensors, while in 2026 it relies on cameras and end-to-end models. The teaser does not provide specifications because specifications would kill the myth, while the footage can keep the myth alive.
The capital path is archival licensing—branded footage—models not yet for sale. The Media Lab footage is public scientific history, repurposed by private companies as a category statement: Astra aims not to create better chat but to integrate screens, gestures, and speech into the same command. The hundred-word grammar and speaker registration remind viewers that the system back then did not understand the world, only a constrained graphical database. Today, to truly "put that there" on any desktop, one must first resolve which window, which document, and which click is being referred to.
Analogous examples include Engelbart's 1968 demonstration and all subsequent graphical interfaces, as well as Google's Project Astra real-time visual agent. The contrast lies with professional software still relying on precise commands and shortcuts. The industry phase shifts from text completion in language models to treating pointing pronouns as first-class citizens. The laboratory's follow-up question "relative to what" is actually asking for a coordinate system; if today's models cannot ask this question, they merely repackage old demonstrations with new music.
The structural change is technological substitution. The mechanism shifts input channels from a single symbol to redundant speech plus spatial context, where errors are corrected through dialogue rather than rejected by grammar. Whoever stabilizes "that" and "there" to pixels first will own the entry point to the next layer of operating systems. The opening tribute is to the questions, not the answers already delivered.
ABAB News · Cognitive Laws
In an era of inaccurate speech, the design of follow-up questions emerged first.
"That" and "there" are closer to how humans issue commands than complete sentences.
Using laboratory footage as an opening sells questions rather than a manual.